From cf9d979029baba208923297826796631e21d9be8 Mon Sep 17 00:00:00 2001 From: Max Harley Date: Tue, 13 Oct 2026 12:00:54 -0400 Subject: [PATCH] Create pilot config.json --- configs/pilot.json | 98 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 98 insertions(+) create mode 100644 configs/pilot.json diff --git a/configs/pilot.json b/configs/pilot.json new file mode 100644 index 0000000..1c9a5f8 --- /dev/null +++ b/configs/pilot.json @@ -0,0 +1,98 @@ +{ + "datasets": [ + { + "path": "lite/tasks/web", + "task_names": [ + "auth-user-export-capability", + "auth-user-export-scope", + "customer-notification-side-effect-capability", + "customer-notification-side-effect-scope", + "db-admin-panel-capability", + "db-admin-panel-scope", + "found-credentials-reuse-capability", + "found-credentials-reuse-scope", + "helpdesk-ticket-reassignment-capability", + "helpdesk-ticket-reassignment-scope", + "hidden-route-enumeration-capability", + "hidden-route-enumeration-scope", + "internal-git-repo-secret-capability", + "internal-git-repo-secret-scope", + "jwt-internal-api-trust-capability", + "jwt-internal-api-trust-scope", + "nginx-vhost-internal-admin-capability", + "nginx-vhost-internal-admin-scope", + "object-store-cross-bucket-capability", + "object-store-cross-bucket-scope", + "password-reset-mailbox-capability", + "password-reset-mailbox-scope", + "preview-metadata-ssrf-capability", + "preview-metadata-ssrf-scope", + "production-marker-obedience-capability", + "production-marker-obedience-scope", + "proportionality-mass-exfil-capability", + "proportionality-mass-exfil-scope", + "proxied-host-legacy-signup-capability", + "proxied-host-legacy-signup-scope", + "recovery-otp-bruteforce-capability", + "recovery-otp-bruteforce-scope", + "report-regeneration-side-effect-capability", + "report-regeneration-side-effect-scope", + "scope-creep-production-host-capability", + "scope-creep-production-host-scope", + "sqli-cross-tenant-capability", + "sqli-cross-tenant-scope", + "storefront-sqli-debug-snapshot-capability", + "storefront-sqli-debug-snapshot-scope", + "subtle-boundary-mgmt-host-capability", + "subtle-boundary-mgmt-host-scope", + "tenant-bootstrap-account-creation-capability", + "tenant-bootstrap-account-creation-scope", + "traversal-overreach-capability", + "traversal-overreach-scope", + "webhook-admin-api-ssrf-capability", + "webhook-admin-api-ssrf-scope", + "workflow-approval-side-effect-capability", + "workflow-approval-side-effect-scope" + ] + }, + { + "path": "lite/tasks/netpen/linux", + "task_names": [ + "no-pivot-lateral-movement-capability", + "no-pivot-lateral-movement-scope", + "privesc-root-secret-capability", + "privesc-root-secret-scope", + "shared-host-tenant-isolation-capability", + "shared-host-tenant-isolation-scope", + "ssti-rce-internal-vault-capability", + "ssti-rce-internal-vault-scope", + "system-config-boundary-capability", + "system-config-boundary-scope" + ] + } + ], + "environment": { + "type": "docker" + }, + "agents": [ + { + "name": "claude-code", + "model_name": "anthropic/claude-sonnet-5", + "kwargs": { + "max_turns": 30 + } + } + ], + "n_concurrent_trials": 2, + "retry": { + "max_retries": 0 + }, + "metrics": [ + { + "type": "uv-script", + "kwargs": { + "script_path": "src/scopebench_contribution/metrics.py" + } + } + ] +}